Synonymy, Lexical Fields, and Grammatical Constructions. A study

Synonymy, Lexical Fields, and Grammatical
Constructions. A study in usage-based Cognitive
Semantics
Dylan Glynn
1. Introduction
tD
in
Pr
e-
Pr
Cognitive Linguistics is, by definition, a usage-based approach to language.
Its model of language places usage at the very foundations of linguistic
structure with a linguistic sign, the form-meaning pair, argued to become
entrenched through repeated successful use. It is this entrenchment that
renders symbolic gestures linguistic rather than merely incidental and represents the key to structure in language. Patterns of language usage across
many individuals can be argued to be indices of shared entrenchment.
When large numbers of language users possess the same or similar entrenchment, we can talk about grammar, that is, linguistic structure.
Importantly, as cognitive linguists, we believe this structure to be conceptually motivated. A basic phenomenon in conceptual structuring is salience. This concerns the conceptual prominence of perceived (or conceived)
objects and their relations. Although frequency represents an important
factor in determining salience, a one-to-one relationship between relative
frequency and relative salience does not exist. Various cultural and perceptual factors can make relatively infrequent concepts salient and vice versa.
Corpus-driven linguistics is frequency based and so inherently restricted in
what it can say about conceptual salience. Nevertheless, frequency data are
perfectly placed to allow us to make generalisations about patterns of usage
across speech communities. Importantly, from a Cognitive Linguistics perspective, we can make the assumption that these patterns of usage represent
speakers’ knowledge of their language, including the conceptual structures
that motivate language. In this indirect way, the inductive generalisations
based on frequency permit us to make hypotheses about the conceptual
structure of language. This is possible without making more theoretically
tenuous claims about the relation of frequency to cognition, such as those
presented in Gries (1999) and Schmid (2000).1
ft
ra
Glynn, D. 2010. Synonymy, Lexical Fields, and Grammatical Constructions. A study in usage-based
Cognitive Semantics. In Cognitive Foundations of Linguistic Usage-Patterns, H.-J. Schmid & S. Handl
(eds), 89-118. Berlin: Mouton de Gruyter.
2 Synonymy, Lexical Fields, and Grammatical Constructions
tD
in
Pr
e-
Pr
This study examines new usage-based techniques to capture semantic
relations between near-synonymous words. The conceptual space encoded
by a language is divided up in complex ways by lexical semantics. It follows that the study of lexical synonymy has a long tradition within Cognitive Linguistics. Moreover, the tradition dates back to some of the first
corpus-driven research within the cognitive framework. Beginning with
Dirven et al. (1982), Lehrer (1982), Schmid (1993), Geeraerts, Grondelaers
and Bakema (1994), and Rudzka-Ostyn (1995) a strong line of empirical
research developed. The current state of the art divides into the study of
lexical near-synonyms (Newman & Rice 2004a, 2004b, Divjak 2006,
Divjak & Gries 2006) and syntactic alternations (Gries 1999, Heylen 2005,
Grondelaers et al. 2007, Speelman & Geeraerts forthc.).2 This study advances upon previous approaches by applying a different statistical technique
and by experimenting with direct semantic analysis in the annotation.
Within Cognitive Linguistics, the use of corpora and empirical methods
more generally represents an important movement. Indeed, many argue that
such approaches are crucial to the advancement of the field (Geeraerts
2006, Gibbs 2007, Croft 2008). The application of such methods to the
study of semantics is not, however, straightforward. Corpus linguistics is
essentially the analysis of large numbers of examples. A corpus linguist
must examine many hundreds or even thousands of utterances before he or
she can make any generalisations. It must be remembered that those generalisations are only valid to the extent that the analysis of those examples is
valid. It is a common myth that corpus linguistics replaces linguistic analysis with quantitative deductions. Nothing is further from the truth. The annotation of a dataset is the laborious linguistic analysis of examples. Often
computational techniques allow one to automate much of that analysis, but
in the field of semantics, this is not possible. This study is concerned with
precisely these quantitative usage-based methods for semantic description
and so annotation is entirely made up of manual semantic analysis.
ra
2. BOTHER: Lexical Field, Conceptual Space, Three Near-Synonyms
ft
2.1. Near-Synonymy and Grammatical Constructions
Synonymy, or more precisely near-synonymy, is the study of semantic
relations between lexemes or constructions that possess a similar usage. In
this study, we focus on three lexemes denoting the concept BOTHER; these
Dylan Glynn 3
are annoy, bother, and hassle. Example (1) captures the kind of semantic
relations in question. We seek to explain speaker choice between these
lexemes.
(1)
People need paypal.... Too much hassle over cheques, especialy
when you cant be bothered to check your statement, god she
annnoyed me.2
tD
in
Pr
e-
Pr
Closely related lexemes have a special place in Cognitive Linguistics because their use, both in terms of their overlap and difference, can be seen as
a reflection of the conceptual structures that motivate language use, and
thus its structure. Although there is a certain circularity in this reasoning,
we can justify approaching the question in such terms because speakers
choose between linguistic forms when they speak. If we assume that speakers have knowledge of their language and culture and make their judgements based on that knowledge, this entails that their choices will reflect
such knowledge. In Cognitive Linguistics, where entrenched language
structure (or knowledge of language use) equates conceptual structure, by
identifying the patterns of similar and distinctive usage, we chart the conceptual structure that motivates those patterns.
The principle is the same for the study of polysemy. Indeed, the cognitive study of polysemy and near-synonymy can be seen as a re-working of
the Structuralist semasiological - onomasiological distinction (see Geeraerts, Grondelaers and Bakema 1994). Seen in this light, polysemy, or
semasiological variation, is the study of the different uses of a form and
synonymy, or onomasiological variation, is the study of the choice between
different forms. If we make generalisations about usage based on large
numbers of examples, then we have a usage-based approach to conceptual
structure. This, of course, must be presented with the caveat that we cannot
make clear deductions about conceptual categorisation and prototypicality
until the relationship between ontological salience and frequency of use is
better understood.
However, it is too simplistic to speak of choices between words. Just as
lexical choices are reflections of different construals, so too are their
grammatical expression. The belief that different ‘lexicogrammatical framings’ or ‘configurational structurings’ that result from the integration of
lexical semantics and different parts or speech and morpho-syntactic forms
represents a fundamental tenet of Cognitive Linguistics (Fillmore 1977:
128, Langacker 1987: 138ff, Talmy 1988: 173ff, Fillmore 2003: 250f).
ft
ra
4 Synonymy, Lexical Fields, and Grammatical Constructions
tD
in
Pr
e-
Pr
When a speaker wishes to express the concept of BOTHER, for instance, it is
unlikely that the speaker decides beforehand and independently of the context that this concept will be profiled nominally or verbally, just as it is
unlikely that, given a verbal choice, he or she will have a predetermined
selection between encoding the concept as an intransitive or transitive
event. The ability to construe events and things, of even the most concrete
nature, means that it will be rare that the speaker has no choice in this matter. If we can assume that the kinds of grammatical semantics associated
with grammatical class and grammatical construction are part of the semantics expressed by the speaker, then they are an integral part of the lexeme
chosen. It is for this reason that we cannot consider only verbs or only
nouns in the study of synonymy.
There are two points to consider here. Firstly, grammatical semantics
are not predictable “additions” to the lexical semantics. Although often the
grammatical profiling of a lexical concept results in regular semantic integration, that is not always the case (Glynn 2002, 2005, 2007, forthc.).
Therefore, we need to treat the interaction between the different grammatical profilings of the lexical concept as onomasiological choices, that is, part
of the synonymous field. Secondly, there is growing evidence that language
knowledge is largely redundant and that speakers rote-learn large amounts
of profiling variation as entrenched units (Dąbrowska 2006). This means,
for example, the simple and the continuous form of a verb or the nominative and instrumental case of a noun are entrenched as separate linguistic
units and not ‘generated by the grammar’. This is in line with Croft's (2001)
arguments for a fundamental Construction Grammar approach to language
structure. For these two reasons, the semantic unpredictability of lexicalgrammatical composition and the fact that many of these compositions are
entrenched as separate form-meanings pairs, if we are to produce a cognitively realistic grammar of lexical choice, we cannot restrict ourselves to
one part of speech. Since from a Construction Grammar point of view,
parts of speech are merely a subtype of grammatical constructions, we will
refer to this formal variation as grammatical class and assume there is only
a theoretical divide between the formal variation of grammatical class and
grammatical construction.3
There is one last complication that must be taken on board in a usagebased approach to synonymy. Since generalising about the entrenched
usage of many individuals is the basis of our grammar, we must account for
variation between those individuals and within that usage. Therefore, Cognitive Semantic study, as a usage-based approach, must necessarily include
ft
ra
Dylan Glynn 5
tD
in
Pr
e-
Pr
what is traditionally considered extralinguistic and social parameters, such
as register and dialect.4 By including this information, we achieve a truly
usage-based description of usage patterns relative to a range of factors such
as age, sex, region, language mode, and register.
We can conclude that the study of lexical near-synonymy is important
and informative from a Cognitive Linguistic perspective since it offers us
an indirect method for mapping conceptual structure via lexical choices.
However, these lexical choices interact in a complex way with formal variation and the grammatical profiling of those lexical concepts. We need,
therefore, to treat near-synonymy across the various grammatical classes
and grammatical constructions that combine with lexical concepts. Lastly,
choice between these forms is made in the social context of their use. Variation between language users and speech contexts surely affects lexical
choice and so these dimensions must also be added to the equation.
We are, therefore, confronted with an inherently multidimensional object of study. We must identify patterns in usage relative to a wide range of
forms and relative to a wide range of contexts. It is this multidimensional
element of language structure that calls for the use of multifactorial statistical techniques to help identify usage patterns. This aspect of usage is not so
readily accessible employing intuitive methods of analysis. Indeed, the
multidimensional element of language structure is not identifiable when
one considers the frequency of the different factors of usage individually.
We need to access the simultaneous interaction of the different factors of
language and to do so we need multifactorial techniques. This study demonstrates why such an approach is necessary and considers one simple
technique for its application. In contrast with previous quantitative studies
of synonymy, which have employed Hierarchical Cluster Analysis
(Divjak 2006, Divjak & Gries 2006), we employ a technique not previously
used for such purposes. This technique, Correspondence Analysis, has the
advantage that it maps correlations rather than simply grouping variables. It
has, however, the disadvantage that its visualisations can be difficult to
interpret.
ft
ra
2.2. Data and Analysis
The data for this study comes from a large non-commercial corpus built
from on-line personal diaries. The language is informal and in many ways
similar to spoken mode. In part, this is due to the “Dear Diary” writing
tradition that involves talking ‘to your diary’, but it is also because these
6 Synonymy, Lexical Fields, and Grammatical Constructions
tD
in
Pr
e-
Pr
diaries ‘speak back’; the LiveJournal on-line diary service used to build the
corpus is interactional. This service allows the readers to respond to the
“blog” entries and they regularly do. Indeed, the authors expect it and they
often complain when their readers do not enter into dialogue. The corpus is
made up of diary entries proper, not the dialogues, but the monologicdialogic distinction is blurred since the writer is assuming that people will
respond to his or her text. Evidence of this may be found in the countless
references to certain readers and frequent switching to second person, both
singular and plural. This results in quite a unique discourse style that is at
once narrative and dialogic.
Despite the richness of the language in its naturalness, the corpus represents only a single text type. This is a basic and inherent limitation for this
study. Corpus representativity is an important and often under estimated
issue for usage-based approaches to language. One must be careful not to
draw conclusions about language based on a single corpus, but at most
about the language type represented in that corpus. For our purposes, the
fact that we consider lexemes that differ in register but we have only one
text type, which is of a most informal nature, is a serious shortcoming.
However, one of the advantages of corpus driven research is that a study
may be repeated on a second corpus and the results compared. For the current purposes, which are to demonstrate the viability and usefulness of the
method, the on-line diary corpus suffices. Needless to say, further research
will be necessary to confirm the results. This is true for both the need of
confirmatory statistical analysis as well as verification through repeat analysis on different data.
From this corpus a relatively even number of the three lexemes were extracted, each with considerable context, totalling approximately 2,000 observations. Across these examples, the proportion of the different parts of
speech, or grammatical classes, for each lexeme is maintained as it occurs
in the corpus. The kind of formal variation in question is best described by
way of example. Examples (2a) - (2h) summarise each of the major classconstruction formal variants in question and serve to introduce the kind of
language that is typical of the corpus.
ra
a.
Saw quite a few people I knew, including the awful stalker
guy who's been hassling me ... (Transitive)
b.
hassle me, bother me, bug me, give me a bad time, If you hassle me about my kinky hair, I'll cut it all off. hat in hand,
humble, almost begging. (Transitive Oblique)
ft
(2)
Dylan Glynn 7
c.
Officer McCoy, me and him was hassling and my gun went
off, hitting him somewhere in his chest. (Intransitive)
d.
thats the LAST time i use a non-digital camera when i'm doing
serious photography because it saves all that ammoying hassle
of SOD'S-BLOODY-LAW!!!!! (Nominal Mass)
e.
I rarely paint my nails(It can be such a hassle!) (Nominal
Count)
f.
It's a very hassily event to do. I believe alot of reasons is it
takes so much time, specially preperation. (Adjective Attributive)
Pr
She will not take part in Saturday's 5000m race, saying she is
tired and bothered (Adjective Predicative)
h.
However, we didn't have the time or the technical know-how
to do this sort of hassling as the PDAs were ordered and the
students were being briefed (Gerund)
tD
in
Pr
e-
g.
ft
ra
Almost all the forms presented here subdivide into further formal variants,
with different syntactic patterns for the verbal forms, grammatical number
amongst the nouns, suffixation for the adjectives, as well as two gerund
forms, one that maintains a verbal argument structure and another that adopts the nominal argument structure. However, these examples represent
the overall pattern of formal variation. Table 1 summarises the relative
number of occurrences of these grammatical classes and constructions.
Altogether some 16 different basic grammatical classes and constructions are found across the three lexemes in the dataset. The eight types
given in Table 1 are the most important numerically and for the practical
concern of data sparseness, the study is restricted to these forms.
8 Synonymy, Lexical Fields, and Grammatical Constructions
Table 1. Principle Classes and Constructions of the Lexical Field BOTHER
Form
Count Noun hassle
Mass Noun hassle
Gerund hassle
Predicative Adjective bother
Intransitive bother
Transitive annoy
Transitive hassle
Transitive bother
Dataset Occurrences
146
217
40
124
222
449
274
275
tD
in
Pr
e-
Pr
The occurrences are annotated for a range of formal, semantic, and extralinguistic features. In total, some 120 features belonging to some 20 partially overlapping variables were analysed and tagged manually. At this
level of onomasiological granularity and with only 2,000 occurrences, the
formal variation in tense, aspect, mood, and post-predicate constituents did
not reveal any informative variation in usage. There was some variation
relative to person and number, but this was found to be an indirect result of
other factors that we examine below. The nature of the corpus limits the
range of extralinguistic variation that may be investigated. For this reason,
the most insightful extralinguistic variable available for consideration is
certainly the regional variation between American and British usage. This
is stratified in the corpus and so straightforward to annotate. For the analysis of the synonymy per se, the semantic variables were the most informative and we will focus on these. Before we examine the variables in question, an important aside should be made.
Within corpus linguistics, there is a very reasonable tendency to avoid
semantic feature analysis. This is for two reasons. Firstly, semantic annotation is largely manual. Such annotation entails a labour and time intensive
process that limits considerably the number of observations that can be
analysed and tagged. Since data sparseness is an ever-present problem in
quantitative studies, this represents an inherent weakness that one wishes to
avoid. Secondly, corpus linguistics, like all empirical methods, seeks to
maximise objectivity. Semantic feature analysis is inherently subjective.
There are strong counterweights to these arguments. Although we can
describe a great deal of linguistic structure limiting our research to formal
phenomena, ultimately, especially within a framework such as Cognitive
Linguistics, we must also apply these kinds of techniques to semantic struc-
ft
ra
Dylan Glynn 9
tD
in
Pr
e-
Pr
ture. Although this will force us to work with smaller numbers of observations, it represents an inherent weakness of the method and it must be taken
on board and considered when we estimate the value of the results it produces.
The same is true for the question of objectivity. We cannot pretend that
any semantic analysis will be purely objective, but this should not stop us
from investigating semantic structure. Quantitative studies of linguistic
semantics simply repeat the kind of semantic analysis that traditional linguists use, but many hundreds of times. Although, in itself, this does not
assure a higher degree of objectivity, the large number of examples does
improve analytical reliability in a number of ways.
Firstly, by examining many hundreds, or thousands, of examples the researcher sees facets of usage that would not necessarily be found through
hermeneutic reflection. Although this approach cannot hope to account for
all possible uses, the analysis of large numbers of found examples offers
the researcher an 'external', therefore objective, source for his or her analysis. However, this does not mean the analysis itself is more objective. Secondly, a quantitative and usage-based approach offers three means for result verification, which serve as check on the objectiveness of the analysis.
In the first place, systematicity and intuitively sound patterns found by the
statistical results are indications of accuracy in semantic analysis. It must
be remembered that after the analysis, the results found through the statistical treatment of the data are independent of the researcher, and in this, are
completely objective. When patterns of usage that match an intuitively
sound perception of usage ‘fall out’ from the statistical analysis, we can be
reasonably sure that the original semantic analysis is accurate. In a second
place, confirmatory statistical techniques employ models of the data, based
on the results of the analysis, to check their validity. If one may predict the
usage of a word, in a given situation, to a very high level of accuracy, then
we can be more sure that the original analysis is accurate. In a third place,
one may repeat the analysis on a second dataset. If the results are comparable, then once again, we can be surer of the accuracy of the semantic
analysis.
We concentrate on three semantic variables, the cause of the BOTHER
event, the affect upon the patient of the event, and the presence or lack of
humour in the description of the event. The annotation focuses not on the
word, but on the entire utterance. In many cases, a great deal of context
needs to be considered to accurately ascertain the cause or affect being
described by the lexeme in question. Table 2 lists the three semantic variables and the features for which they are annotated.
ft
ra
10 Synonymy, Lexical Fields, and Grammatical Constructions
Table 2. Semantic Features
Cause of Event
expenditure of energy
imposition
imposition / request
request
interruption
condemnation
tease
aesthetics
repetition
Affect on Patient
anger
concern - thought
emotional pain
physical pain
Humour
Presence of humour
Absence of humour
tD
in
Pr
e-
Pr
In order to avoid overlap between the variables, either the cause or the
affect was coded, never both. Statistical techniques do not work when one
has redundancy across variables. Certain cause features, for example, ‘repetition’, which systematically co-occurs with what would be the affect of
‘boredom’, are therefore a problem. Thus, for the purposes of the statistical
analyses below, the cause and affect variables are treated as a single variable.
Most of the features should be self-explanatory, however several warrant a word of explanation. Three particularly important features include
‘imposition’, ‘imposition-request’, and ‘request’. These features identify
uses where the agent of the event imposes him or herself upon the patient
or makes a request of him or her. Often, both these two features are present;
when this is the case, the example is coded as ‘imposition-request’. The
clearest way to explain these features is by way of example. Examples (3a)
- (3c) represent these semantic distinctions.
(3)
While Valentine's Day is a nice thought, it's always such a
hassle. Romance should never be an obligation, and neither
should it be restricted to a single day, which are the messages Valentine's Day sends. (Imposition)
b.
... and walked up the Grays Inn Road being hassled by aggressive beggars who glared at me straight in the eyes, asking Got any change? (Imposition request)
c.
I can then update the page, and won't need to hassle you for
the results of matches that have been postponed. (Request)
ft
ra
a.
The features ‘aesthetics’, ‘condemnation’, and ‘tease’ also deserve explanation. In the diary entries, speakers often experience BOTHER because
Dylan Glynn 11
someone is judging them. This is quite distinct from a situation where
classmates or friends are teasing the patient and also from a situation where
some inherent quality in the world displeases the patient. Again, examples
can clarify the semantic features in question as well as the kind of subtle
semantic differences that the coding seeks to capture. A reasonably large
amount of context was needed in order to accurately discern many of the
semantic distinctions.
(4)
"Now it's tough being an American. Everyone always gives
us hassle for having a stupid president. Especially you
Brits. You give us hassle for having a retard for a President.
But we know he's a retard. (Condemnation)
b.
bumping into Kath, which i always do when i'm fucked,
and having lots of hugs. and not being able to pee in front
of her in the toilets and hassling her because she has curly
hair and i wanted to "ping" it. (Tease)
tD
in
Pr
e-
Pr
a.
c.
he dnt reilise tht she loves him sooo much it dnt bother her
wot is on his face lol (Aesthetic)
It should also be stressed that ‘humour’ refers to the utterance in which the
lexeme is used and to the intention of the speaker. The other features
should be self-explanatory, their semantic distinctions being drawn in a
similar manner to those described here.
3. Usage-Based Methodology. A Multifactorial Treatment of Results
3.1. Semantic Relations between Lemmata
ft
ra
Having completed the semantic analysis of the observations, we now have
what are referred to as multiway contingency tables. These are three, fourway, or n-way tables of frequencies of co-occurring, extralinguistic, formal,
and semantic features. Although one may not visualise a multiway table,
the mathematical relations are simply the frequencies of co-occurrences of
multiple features. These features are relative to various levels of granularity
in the formal variation. For example, we can examine the correlation between the semantic variables and the three words without including the
formal variation of each lemma. We can equally zoom in and examine the
formal variation at a very fine-grained level, differentiating not only grammatical class and grammatical construction but also tense, mood, aspect,
12 Synonymy, Lexical Fields, and Grammatical Constructions
tD
in
Pr
e-
Pr
and so forth. The limitation is data sparseness: as we include more detail in
formal variation, the numbers of occurrences for each semantic feature
drops quickly. At a certain point the frequencies of occurrences become too
small for us to identify meaningful generalisations in the data.
Moreover, interpreting a three-way or four-way table of frequencies of
co-occurrences is not possible without using multivariate tools. Exploratory
techniques exist that search through these tables looking for patterns of
correlations. In other words, mathematically, some features co-occur appreciably more often than others. In our case, these are the semantic features co-occurring with the various forms of annoy, bother, and hassle. One
such exploratory technique is Correspondence Analysis. This simple statistical technique takes the frequencies of multiway tables and converts those
frequencies to distances. It then conflates the multidimensional distances to
a two-dimensional plane that maps the correlations between the features
visually. Although this allows us to ‘see’ the correlations and differences
between the forms and semantic features, one must be careful in reading
such visualisations since, obviously, representing n-dimensions in a twodimensional plane can be misleading. For this reason, the position of many
of the data points relative to other data points can be misleading. Careful
consultation and experience interpreting the plots is the only way to avoid
misinterpretation.
Let us begin with a Bivariate Correspondence Analysis of the semantic
variables relative to the three lemmata. Figure 1 is a correspondence map of
the analysis. It should be remembered that relative proximity of the data
points represents relative correlation.
ft
ra
Dylan Glynn 13
(iii)
(i)
tD
in
Pr
e-
Pr
(ii)
Figure 1. Correspondence Analysis BOTHER Lemmata and Cause-Affect
ft
ra
Interpreting the visualisations of Correspondence Analysis can be difficult. Let us move through a description of the plot, step by step. Firstly, on
the left (i), we see annoy, grouped with ‘anger’ <ang> and ‘interruption’
<int>. The feature ‘anger’ <ang> is to the left of annoy, which stands between it and the other lemmata. The position of this feature shows that it is
highly distinctive for the usage of the lemma annoy. This is intuitively
sound: of the three lexemes in question, annoy represents the point of overlap with the concept of ANGER, an interpretation corroborated by traditional
dictionaries. Also associated with the lemma annoy is ‘interruption’ <int>.
However, the fact that this feature occurs to the right of the annoy data
point, placed between the two other constructions, suggests that despite a
clear association with annoy, this feature is shared to some extent by all
three words.
Placed more or less evenly between (i) annoy and (ii) bother, we find
two cause features, ‘aesthetics’ <aesth> and ‘repetition’ <rep>. We can
suppose quite safely that these two features are characteristic of both these
lemmata. The two features ‘concern – thought’ <thght> and ‘emotional
pain’ <pain> lie just beneath the bother data point and so are distinctly
associated with this lemma. Just as ‘anger’ is effectively unique to annoy,
14 Synonymy, Lexical Fields, and Grammatical Constructions
tD
in
Pr
e-
Pr
the semantically similar features ‘emotional pain’ and ‘concern – thought’
are effectively unique to bother. This is also intuitively sound. A third feature, which was rare in the data, is also highly associated with the lemma
bother. The cause feature ‘physical pain’ <phys> only occurs 10 times out
of almost 2,000 observations. Of these 10 occurrences, 7 are with bother, 2
with hassle, and 1 with annoy. It seems with such small frequencies, we
cannot draw any firm conclusions. However, in the dataset, to the extent
that this feature occurs, it is associated with bother.
One of the three most important features in terms of frequency, occurring 650 times, is that of the ‘expenditure of energy’ <engy>. Its data point
lies in the centre of the plot, equidistant from hassle and bother, yet relatively far from annoy. The position of this data point strongly suggests that
this feature is characteristic of bother and hassle, more than of annoy.
Finally, the cluster in the top right (iii) sees hassle associated with a
large number of overlapping semantic features. One feature, ‘imposition’
<imp>, is distinct from this micro-cluster and considerably closer to the
data point of hassle. This may signify a stronger correlation but needs further verification. The dense cluster just above this point consists of request
<req>, ‘imposition request’ <imp_req>, ‘condemnation’ <condemn>, and
‘tease’ <tease>. These four semantic features seem to identify two ‘meanings’ of the word, the ‘imposition request’ and simple ‘request’ features
being semantically similar as well as the ‘tease’ and ‘condemnation’ features clearly carving out a similar semantic space.
We could not ask for clearer results in this first Correspondence Analysis . Each of the three lemmata are evenly dispersed across the plot, distinctly grouped by semantic features. Certain semantic features lie between
the lemmata, showing overlap in the semasiological distribution. This kind
of semantic map is a simple but powerful generalisation that shows the
basic differences and similarities of usage across the three synonymous
words.
At this point, it is worth noting that mapping the correlations between
such semantic features and various forms should be seen as an indirect
means for capturing the conceptual structure. The kind of the results we see
here are intuitively sound and match the kind of results that one would posit
using an individual’s knowledge of a language. The important difference,
of course, is that this technique permits repeat analysis, and is therefore
easily verifiable.
ft
ra
Dylan Glynn 15
annoy
hassle
anger
interruption
energy
repetition
aesthetics
bother
imposition
request
request imposition
condemn
agitate
tease
thought-concern
emotional pain
physical pain
Figure 2. Box Summary BOTHER. Lemmata and Cause-Affect Features
Pr
tD
in
Pr
e-
We can summarise the results of the Correspondence Analysis with a box
diagram. This is presented above in Figure 2. Although the box diagram
adds nothing to the actual results, it is clear and more easily interpretable.
Its downside is, by rendering the correlations discrete, it does not capture
the semantic continua between the correlations.
Despite these intuitively attractive results, even dictionaries break down
lemmata into grammatical classes and this kind of coarse-grain analysis is
only helpful in mapping the aggregate meaning of the three words. Any
accurate semantic description must look closer than this.
3.2. Grammatical Class, Grammatical Construction, and Semantic
Similarity
ft
ra
Let us now repeat the analysis while rendering the formal dimension more
fine-grained. Figure 3 plots a Correspondence Analysis that identifies correlation between cause-affect and class-construction.
In direct contrast to the lemma level of analysis, we see more semantic
similarity between different words within the same class-construction than
between the different forms of a single lemma. In group (i), we see how,
relative to the semantic features in question, the transitive forms of annoy
and bother group together. In contrast to this, the transitive use of hassle
sees a distinct usage (ii), highly associated with instances of impositions
and requests. Then a third group (iii) clusters the adjectival, nominal, and
intransitive profilings of all three words.
16 Synonymy, Lexical Fields, and Grammatical Constructions
bother_trans
(iii)
(i)
tD
in
Pr
e-
Pr
(ii)
hassle_trans
Figure 3. Correspondence Analysis of BOTHER. Class-Construction and CauseAffect
ft
ra
Before we look more closely at the detail of these correlations, let us add
another dimension to the analysis. Regional variation often has a profound
effect on semantic variation. This is because even if a word or construction
exists across all the varieties of a given language, this does not entail that it
is used in the same manner. The countless ‘false friends’ between British
and American English are testimony to this. However, if we distinguish the
forms further, dividing between the British and American varieties, the
analysis reveals an almost identical picture suggesting at this onomasiological level, there is little dialect variation. The plot in figure 4 visualises
a Bivariate Correspondence Analysis of class-construction distinguished for
dialect, correlated with the semantic features of cause and affect.
By splitting the class-constructions into British and American variants
we double the number of forms, leading to a denser plot. Moreover, splitting the data offers two datasets for comparison. Assuming there is no substantial dialect variation, this serves as an indirect way of verifying the
results. In light of this, the most important result of the Correspondence
Analysis visualised in Figure 4 is that the three basic uses across the onomasiological field are maintained. Indeed in terms of placement and proximity, the map is little different to that given by the Correspondence Analy-
Dylan Glynn 17
sis of the formal variation without the variable of dialect. The greatest difference in the results is that the outlying cause-affect features, with the
exception of ‘imposition’ <imp>, have been ’brought into’ the clusters. In
the majority of cases, the dialectical pairs behave in the same manner. Only
one pair splits between the different clusters; the Adjectival Construction
for bother.
(i)
tD
in
Pr
e-
Pr
(ii)
(iii)
Figure 4. Correspondence Analysis of BOTHER. Class-Construction-Dialect and
Cause-Affect
Let us look again, this time more closely, at the clusters. We can zoom in
on each of the clusters identified in Figure 4 to see what features and forms
are correlated.
ft
ra
18 Synonymy, Lexical Fields, and Grammatical Constructions
Usage Cluster 1
Dialect Class Form
a.
Transitive annoy
Transitive bother
b.
Am. Predicative bother
(i)
Pr
Affect Features
a.
anger
thought-concern
emotional pain
physical pain
b.
tD
in
Pr
e-
repetition
interruption
aesthetics
ft
ra
The most surprising result here is that the American predicative form of
bother has been clustered with these transitive forms. By dividing the
words into two dialectally distinguished forms, we substantially reduce the
number of co-occurrences with the various semantic features. This may
mean that for a relatively infrequent form such as the predicative bother,
the results are erroneous. We will assume the accuracy of the correspondence analysis, but in this case, further investigation is necessary.
The two transitive forms of bother and annoy cluster with what seem to
be two sets of similar semantic features. Firstly, there appears to be a semantic cline from the affect of ‘anger’ through ‘emotional pain’ and
‘thought-concern’ to perhaps ‘physical pain’. The similarity of these semantic features suggests a clear ’meaning’ is associated with these two
forms. Moreover, the systematicity represented by the grouping of these
semantic features adds weight to the argument that the analysis and annotation has successfully operationalised the subjective nature of these features.
The second sub-group of semantic features found here is less homogenous, but still reasonably coherent. This group, in contrast to the other features, includes causes that are of a relatively inconsequent nature. Causes
such as ‘repetition’, ‘interruption’, or ‘aesthetic displeasure’ are similar in
that they are little more than inconveniences for the patient.
Dylan Glynn 19
The kind of usage in question can be explained by way of example. The
‘anger’, ‘thought-concern’, and ‘emotional pain’ uses of the transitive annoy and transitive bother are represented by examples (5a) – (5c). This is
contrasted by examples (6a) – (6b), which are typical of causes such as
‘interruption’ and ‘aesthetic displeasure’.
(5)
There are even people out there that annoy the hell out of
me. (Anger)
b.
they can get 2 fuk.. im not gona let it bother me.. (Thoughtconcern)
c.
It bothers me when I am starting to beg for people to think
about me when I've never done it before. I cannot explain
how I feel right now. (Emotional pain)
Pr
a.
(6)
a.
tD
in
Pr
e-
oh on the last night the guys kept annoying him while he
was trying to sleep (Interuption)
b. Ok, I don't really like my mood theme. I love Nightmare
and all but the theme is bothering me for some reason.
(Aesthetic)
Usage Cluster 2
Dialect Class Form
Brit. Transitive hassle
Am. Transitive hassle
Affect Features
a.
condemnation
tease
b.
ra
imposition
request
imposition-request
ft
Here, we see that the transitive form of hassle stands out as a relatively
unique usage. It is associated with two very clearly grouped sets of semantic features. Again the systematicity of the semantic feature groupings
strongly supports the success of this variable’s analysis and annotation.
These groups include, on the one hand, ‘tease’ – ‘condemnation’ and on the
other hand, ‘imposition’ – ‘request’ – ‘imposition-request’. It seems that
20 Synonymy, Lexical Fields, and Grammatical Constructions
this form is distinct in its usage and possesses two relatively distinct meanings. Examples (7) and (8) represent the two clusters of features and the
kind of usage they indicate.
(7)
a.
Some smokers also have a dream that someday the nonsmoking world will quit hassling them about their smoking.
(Condemnation)
b.
Anyway today nothing excited happen excpet all my teachers had to be taught to be better teachers (their turn to be
bored) and I had substitutes in every class. I hassled them
and had fun doing it! (Tease)
Pr
(8)
i saw him yesterday and he was being all touchy feely.....i
don't want him back...but hes hassling me now and I fee
sorry for liz(his new g/f) (Imposition)
b.
she had other ideas and hassled Dave to walk her to the
train station (Request)
Ford and Greg: Nah, the real Glasgow neds hassle us for
our wallets (Imposition-request)
tD
in
Pr
e-
a.
c.
Usage Cluster 3
Dialect Class Form
a.
Intransitive bother
b.
Mass hassle
Count hassle
Gerund hassle
c. Brit. Predicative bother
ra
Affect Features
energy
agitation
ft
The third usage cluster of correlations includes the nominal-gerundive
forms as well as the intransitive forms. Before the addition of the variable
of dialect, it also included the adjectival forms. First, it must be noted that a
wide range of forms are grouped relative to only two semantic features,
‘expenditure of energy’ <engy> and ‘agitation’ <agit>. Second, the first of
these two semantic features is the most common of the dataset and the sec-
Dylan Glynn 21
tD
in
Pr
e-
Pr
ond is a relatively infrequent feature. From this, we can tentatively deduce
that in fact the non-verbal forms are associated with the ‘expenditure of
energy’ relative to the verbal forms, which represent a semantically more
complicated profiling. The correlation with the feature of ‘agitation’ is
likely to be incidental.
Lastly, the British form of the predicative remains in this cluster where
it was before we added the variable of dialect. This is in contrast to the
American predicative form, which as we saw, is found in Cluster 1. However, by adding the variable of dialect, we increase the number of correspondences calculated by the analysis considerably. For a relatively infrequent form, such as the predicative bother, we are faced with a degree of
data sparseness. It is therefore possible that the results presented in Figure
3, are misleading. If this were the case, it would leave all the non-verbal
uses together and associate them with the single most common semantic
feature, the ‘expenditure of energy’. Further investigation is needed in
order to determine if there is a distinction in use between the dialects and
whether this adjectival form does, in fact, divide along the lines suggested
by the analysis.
Let us add one last variable, that of ‘humour’. For such negative emotion terms as annoy, bother, and hassle, this feature is clearly marked. It is
important since it captures a difference that further distinguishes one of the
forms, transitive hassle. In Figure 5, the most striking feature is that the
clustering captured by the analysis remains stable after the addition of the
extra variable. This further re-assures us that the analysis is capturing real
semantic structures extant in language use. However, the feature itself
proves to be important. The lack of ‘humour’ <NHum> falls squarely between both the transitive bother - annoy cluster and the nominal-adjectivalintransitive cluster contrasted starkly by the clear correlation between the
presence of ‘humour’ <Hum> and the transitive hassle uses. Example (9)
captures the kind of uses in question.
a.
Vicky spent most of the days hassling cows and sheep. Occasionally she would do a little skip or run for no reason
b.
... sitting outside Mcdonalds and hassling kids for change,
and taxing people. The west end is the Crewe chav centre,
other wise known as "The Cronx".
ft
ra
(9)
22 Synonymy, Lexical Fields, and Grammatical Constructions
tD
in
Pr
e-
Pr
Figure 5. Multiple Correspondence Analysis of BOTHER. Class-ConstructionDialect, Cause-Affect, and Humour
ft
ra
We can perform one last statistical analysis to help verify our findings.
Hierarchical Cluster Analysis functions in a similar way to Correspondence
Analysis, converting frequencies to distances. However, instead of plotting
those distances, it uses a pre-determined distance measure to identify clusters. The visualisation takes the form of a dendogram. This does not show
what semantic features cause the clustering of the forms, but it does offer a
clearer picture and allows us to include significance testing via bootstrap
resampling. Bootstrapping is a complicated mathematical procedure for
determining the probability that a given result will be repeated, given the
same data. In the plot below, the different forms are clustered relative to the
semantic features cause-affect and humour.
The results clearly verify the results of the Correspondence Analysis.
Not only are the same clusters identified, a further more subtle distinction is
added. Although the intransitive forms, adjectival, and nominal-gerund
forms are grouped together, they are once again subdivided.
Dylan Glynn 23
tD
in
Pr
e-
Pr
Figure 6. Hierarchical Cluster Analysis (Ward) BOTHER. Class-Construction
Cause-Affect
ft
ra
In the plot, the boxes drawn around the dendogram clusters are the bootstrapping results. Two different bootstrapping algorithms are used. The
numbers at the top of the boxes represent the results of the bootstrap samples, the first number is the results of the more reliable multiscale bootstrap
sample and the second number the simpler and less reliable normal bootstrap. The closer the figure is to 100, the better the result. In terms of
probability we have excellent results that strongly suggest these clusters are
accurate representations of the data.
Note that the Cluster Analysis identifies a distinction that is not apparent
in the Correspondences Analyses. What was referred to as cluster 3 above,
is here subdivided into two sub-clusters: intransitive bother and mass noun
hassle on the one hand versus gerund hassle, count noun hassle, and adjectival bother on the other. Investigation into this distinction is beyond the
scope of the current study, but the Cluster Analysis suggests that there is a
clear usage difference between these two groups. Most importantly, the
bootstrapping on the Cluster Analysis offers us a means of verification for
the results found in the Correspondence Analysis. It shows that there is an
extremely high probability that if we repeated this study many hundreds of
times, we would obtain the same groupings of form and usage.
24 Synonymy, Lexical Fields, and Grammatical Constructions
Transitive annoy
Transitive bother
Nominal hassle
Gerund hassle
Intransitive bother
Adjectival bother
anger
concern-thought
emotional pain
physical pain
nonhumour
interruption
repetition
aesthetics
energy
agitation
imposition request
request
condemnation
tease
tD
in
Pr
e-
Pr
imposition
Transistive hassle
humour
Figure 7. Box Summary BOTHER. Class-Construction and Cause-Affect and Humour
ft
ra
By way of conclusion, Figure 7 presents a box summary of the findings.
The results, when summarised in this manner, resemble the conceptual
space maps of the Structuralist era. However, the results presented here fall
out from a mathematical logarithm that examines frequencies of cooccurring features of language use. This does not at all prove the results,
indeed far from it, nor does it necessarily mean they are more accurate.
However, it does mean that the analysis is repeatable. This can be done
with similar data from the same corpus to verify that this is indeed an accurate depiction of the semantic structure associated with the three words for
this kind of language. However, this verification can also be performed
with different corpora of different kinds of language to determine to what
degree the results are influenced by the register and mode of the language
rather than the lexical semantic structure per se. These possibilities for
verification are an important addition to Cognitive Semantic analysis, especially since this method can be expanded to more culturally rich concepts.
Despite the fact that the discrete boxes used to summarise the results of
the Correspondence Analyses may be misleading in their simplicity, they
do help appreciate how, via the careful semantic annotation of some 2,000
Dylan Glynn 25
examples, quantitative investigation helps us map semantic structure. The
diagram can be seen as a summary of the conceptual associations of different yet similar linguistic forms. By adding other semantic features, such as
‘agent type’ and ‘patient type’, ‘topic of discourse’, as well as more formal
detail, such as variations in post-predicate argument structure and so forth,
we could enrich this map, adding finer levels of granularity of formal and
semantic detail. For this, perhaps extra examples would be needed since the
more factors one considers simultaneously, the more data one requires.
Nevertheless, this small study has hopefully shown how quantitative techniques can capture semantic similarity between words and do so while accounting for some of the multidimensionality of language.
Pr
4. Summary
tD
in
Pr
e-
This study has successfully made four points. Firstly, we have seen how
quantitative and multidimensional techniques can help map usage patterns,
patterns that theoretically represent the grammar of that language. In this
way, we have seen how we can vary the level of granularity of the study by
increasing the degree of formal details considered, contrasting a study at
the level of the lemma with a study at the level of grammatical class and
construction. Secondly, we have seen how it is possible to use direct semantic analysis in quantitative approaches. The semantic features in question may be determined subjectively, but the systematicity and intuitively
coherent results demonstrate that careful analysis and annotation of even
subjective semantic characteristics of language use is operationalisable.
Thirdly, we saw how a simple statistical technique, Correspondence Analysis, can help capture the multidimensional correlations produced by the
semantic analysis. Although the discussion did not directly compare Correspondence Analysis with other techniques that have been used to describe
synonymous relations, the technique proved successful. Fourthly and returning to the first point, we have seen how the study of synonymy and
semantic relations of similarity can be used to posit hypothetical conceptual
structures. Since we argue that usage is conceptually motivated, the patterns in usage represent more than grammar, but the conceptual structures
argued by Cognitive Linguistics to motivate grammar. Quantitative usagebased studies of this kind, therefore, offer an indirect yet verifiable approach to the study of conceptual structure.
ft
ra
26 Synonymy, Lexical Fields, and Grammatical Constructions
tD
in
Pr
e-
Pr
There are, of course, certain deductions that this study cannot draw.
Firstly, we are in no place to make hypotheses about the categorisation of
the concepts. It may well be that in these instances, the frequency data do
represent prototype effects and category structure, but until we understand
the relationship between ontological salience and frequency, this is an assumption we cannot make. Secondly and similarly, we cannot draw any
conclusions about the cognitive salience and the processing of the lexical
semantics and its integration with the grammatical semantics. At this level,
corpus-driven research must pass the torch to psychological experimentation, for its frequency counts offer few insights.
To the extent that the corpus is representative of language and to the extent that the dataset is representative of the corpus, we can propose a partial
semantic map of the lexical encoding of the concept BOTHER. There are
other words and expressions that should be included, just as different registers and modes of language, and so we cannot say that we have fully described the synonymy of these words or the conceptual structure they are
used to represent. However, we have a partial map of the patterns of language use, patterns we argue indicate conceptual structure.
The next step will be to test these findings. This needs to be done at two
levels. Firstly, new data from a different sample of language need to be
analysed and the results compared. Secondly, confirmatory statistical techniques need to be used to demonstrate that for the datasets in question, the
results are more than chance and do map, or model, the reality of the data.
Perhaps in comparison to other methods of language analysis, these results
seem conditional and limited. Even if this is true, the results are verifiable
and are truly usage-based representations of the linguistic patterns that
make up the grammar of a language.
Notes
3.
Note that both authors have since stepped back from the stronger claims made
in this vein. For more recent discussion on the relationship between frequency
based evidence and cognition, see Glynn (2006, in press), Schmid (2007), and
Gilquin (2008).
All examples are taken from a corpus built from on-line personal diaries. The
details of which are given in section 2.2.
Further discussion concerning these lines of research and the methods used
may be found in Tummers et al. (2005) and Heylen et al. (2008).
ft
2.
ra
1.
Dylan Glynn 27
4.
5.
Glynn (2004, 2009) goes further to argue that lexical study is not at all possible without morpho-syntactic context. It is argued that grammatical semantics are inherently interwoven with lexical semantics and, regardless of redundancy, the only way to explain lexical structure is by simultaneously
accounting for grammatical structure.
The importance of extralinguistic factors in Cognitive Linguistics is gaining
wide acceptance. See Geeraerts (1995), Kristiansen & Dirven (2008), Geeraerts et al. (forthc.) for discussion and examples of this line of research.
References
tD
in
Pr
e-
Pr
Croft, William
2001
Radical Construction Grammar, Syntactic Theory in Typological
Perspective. Oxford: OUP.
Croft, William
2008
Toward a social cognitive linguistics. V. Evans & S. Pourcel (eds),
New Directions in Cognitive Linguistics, 395-420. Amsterdam: Benjamins.
Dąbrowska, Ewa
2006
Low-level schemas or general rules? The role of diminutives in the
acquisition of Polish case inflections. Language Sciences 28: 120135.
Dirven, René. Goossens, Louis. Putseys, Yves. & Vorlat, Emma
1982
The scene of linguistic action and its perspectivization by SPEAK,
TALK, SAY, and TELL. Amsterdam: Benjamins.
Divjak, Dagmar
2006
Delineating and Structuring Near-Synonyms. St. Th. Gries &
A. Stefanowitsch (ed.), Corpora in cognitive linguistics, 19-56. Berlin: Mouton.
Divjak, Dagmar & Gries, Stefan. Th.
2006
Ways of trying in Russian: clustering behavioral profiles. Journal of
Corpus Linguistics and Linguistic Theory 2: 23-60.
Fillmore, Charles
1977
Topics in Lexical Semantics. R. Cole (ed.), Current Issues in Linguistic Theory. 76-138. Bloomington: Indiana University Press.
Fillmore, Charles
2003
Form and Meaning in Language. Standford: CSLI.
Geeraerts, Dirk
2005
Lectal data and empirical variation in Cognitive Linguistics.
F. J. Ruiz de Mendoza Ibañez & S. Peña Cervel (eds), Cognitive
ft
ra
28 Synonymy, Lexical Fields, and Grammatical Constructions
tD
in
Pr
e-
Pr
Linguistics. Internal Dynamics and Interdisci plinary Interactions,
163-189. Berlin: Mouton.
Geeraerts, Dirk
2006
Methodology in Cognitive Linguistics. G. Kristiansen, M. Archard,
R. Dirven, & F. J. Ruiz (eds), Cognitive Linguistics: Current Applications and Future Perspectives. 21-49. Berlin: Mouton.
Geeraerts, Dirk Grondelaers, Stefan. & Bakema, Peter
1994
The Structure of Lexical Variation. Meaning, Naming, and Context.
Berlin: Mouton.
Geeraerts, Dirk, Kristiansen, Gitte, & Peirsman, Yves (eds)
Forthc. Advances in Cognitive Sociolinguistics. Berlin: Mouton
Gibbs, René
2007
Why cognitive linguists should care more about empirical methods.
M. Gonzalez-Marquez I. Mittelberg, S. Coulson, & M. Spivey (eds),
Methods in Cognitive Linguistics, 2-18. Amsterdam: Benjamins.
Gilquin, Gaëtanelle
2008
What you think ain’t what you get: highly polysemous verbs in mind
and language. J.-R. Lapaire, G. Desagulier, & J.-B. Guignard (eds),
From Gram to Mind. Grammar as Cognition, 237-258. Bordeaux:
Presses universitaires de Bordeaux.
Glynn, Dylan
2002
Love and Anger. The grammatical structure of conceptual metaphors. Style. Cognitive Approaches to Metaphor 36:541-559.
Glynn, Dylan
2004
Constructions at the Crossroads. The Place of construction grammar
between field and frame. Annual Review of Cognitive Linguistics
2:197-233.
Glynn, Dylan
2005
Concept Delimitation and Pragmatic Implicature. Issues for the study
of metonymy. K. Kosecki (ed.), Perspectives on Metonymy, 157-174.
Frankfurt: Lang.
Glynn, Dylan
2006
Iconicity and the Grammar - Lexis Interface. E. Tabakowska, C.
Ljungberg & O. Fischer (eds), Iconicity in Language and Literature 5, 267-286. Amsterdam: Benjamins.
Glynn, Dylan
2007
Rain and Snow in West Germanic. A test case for cognitive grammar. J.-R. Lapaire, G. Desagulier, & J.-B. Guignard (eds), From
Gram to Mind: Grammar as Cognition, 191-212. Pessac: Presses
Universitaires de Bordeaux.
Glynn, Dylan.
2008
Arbitrary Structure, Cognitive Grammar, and the partes orationis. A
Study in Polish Paradigms. K. Willems & L. De Cuypere (eds),
ft
ra
Dylan Glynn 29
Naturalness and Iconicity in Linguistics, 215-239. Amsterdam: Benjamins.
tD
in
Pr
e-
Pr
Glynn, Dylan
2009.
Polysemy, Syntax, and Variation. A usage-based method for Cognitive Semantics. V. Evans & S. Pourcel (eds). New Directions in
Cognitive Linguistics, 77-105. Amsterdam: Benjamins.
Glynn, Dylan
Forthc. The semantics of extralinguistic variation. A quantitative study of
dialect effects on polysemy. D. Geeraerts, G. Kristiansen, & Y.
Peirsman (eds), Advances in Cognitive Sociolinguistics. Berlin:
Mouton.
Gries, Stefan. Th.
1999
Particle movement: a cognitive and functional approach. Cognitive
Linguistics 10:105-45.
Grondelaers, Stefan. Speelman, Dirk & Geeraerts, Dirk
2007
A Case for a Cognitive Corpus Linguistics. M. Gonzalez-Marquez I.
Mittelberg, S. Coulson, & M. Spivey (eds), Methods in Cognitive
Linguistics, 149-170. Benjamins: Amsterdam.
Heylen, Kris
2005
A Quantitative Corpus Study of German Word Order Variation.
Stephan Kepser & Marga Reis (eds.), Linguistic Evidence: Empirical, Theoretical and Computational Perspective, 241-264. Berlin:
Mouton.
Heylen, Kris Tummers, José, & Geeraerts, Dirk
2008.
Methodological issues in corpus-based Cognitive Linguistics.
G. Kristiansen & R. Dirven (eds), Cognitive Sociolinguistics: Language Variation, Cultural Models, Social Systems, 91-129. Berlin:
Mouton.
Kristiansen, Gitte & Dirven, René (eds)
2008
Cognitive Sociolinguistics: Language Variation, Cultural Models,
Social Systems. Berlin: Mouton.
Langacker, Ronald
1987
Foundations of Cognitive Grammar. Vol. 1, Theoretical Prerequisites. Stanford: Stanford University Press.
Lehrer, Adrienne
1982
Wine and Conversation. Bloomington: Indiana University Press.
Newman, John & Rice, Sally
2004a
Patterns of usage for English SIT, STAND, and LIE: A cognitivelyinspired exploration in corpus linguistics. Cognitive Linguistics 15:
351-396.
ft
ra
30 Synonymy, Lexical Fields, and Grammatical Constructions
tD
in
Pr
e-
Pr
Newman, John & Rice, Sally
2004b
Aspect in the making: A corpus analysis of English aspect-marking
prepositions. S. Kemmer & M. Achard (ed.), Language, Culture and
Mind, 313-327. Stanford: CSLI.
Rudzka-Ostyn, Brygida
1995
Metaphor, Schema, Invariance: The Case of Verbs of Answering. L.
Goossens, P. Pauwels, B. Rudzka-Ostyn, A.-M. SimonVandenbergen & J. Vanparys (eds), By Word of Mouth. Metaphor,
metonymy and linguistic action in a cognitive perspective, 205-244.
Amsterdam: Benjamins.
Schmid, Hans-Jörg
1993
Cottage, idea, start: Die Kategorisierung als Grundprinzip einer
differenzierten Bedeutungsbeschreibung. Tübingen: Niemeyer.
Schmid, Hans-Jörg
2000
English Abstract Nouns as Conceptual Shells. From Corpus to Cognition. Berlin: Mouton.
Schmid, Hans-Jörg
2007
Entrenchment, Salience, and Basic Levels. D. Geeraerts & H.
Cuyckens (eds), The Oxford Handbook of Cognitive Linguistic, 117138. Oxford: OUP.
Speelman, Dirk & Geeraerts, Dirk
Forthc. Causes for causatives. The case of Dutch doen and laten. Linguistics
of Causality. T. Sanders & E. Sweetser (eds). Cambidge: Cambridge
University Press
Talmy, Leonard
1988
The relation of grammar to cognition. B. Rudzka-Ostyn (ed.), Topics
in cognitive linguistics. Amsterdam: Benajmins.
Tummers, José, Heylen, Kris & Geeraerts, Dirk
2005
Usage-based approaches in Cognitive Linguistics: A technical state
of the art. Corpus Linguistics and Linguistic Theory 1: 225-26.
ft
ra